VCF-42: Reduce memory consumption when a single position returns many samples

[alancleary]

Previously the header for every sample returned for a single position was being loaded into memory. However, many headers are identical except for their sample line, meaning basically the same headers were being loaded multiple times per position.

To alleviate the issue, this PR computes the MD5 checksum for each header (minus the sample line) and only stores the unique ones. Samples are then associated with the correct header using an associative array.

Notes about the implementation:

• HTSlib includes low-level support for computing MD5 checksums, but the only high-level functions provided for computing checksums on user-space data are in the context of the HTSlib CRAM API. Thus this PR includes a minimal external library that provides the requisite high-level API. Rolling the HTSlib low-level API into an adequate high-level API is left as future work.
• MD5 collision detection is implemented but there is no mitigation; a run-time exception will be thrown that provides details about the collision.
• Running on the VCF-42 reproducer query, the query run-time is slightly faster than before and the memory usage was reduced 24x. It's worth noting that disabling collision detection cuts the run-time nearly in half. The ability for users to disable collision detection is currently not implemented.
• Allowing users to disable collision detection and/or choose the checksum algorithm may be features worth implementing in the future.